[lworld] RFR: 8364464: [lworld] GCs do not process flat arrays by chunks [v2]

[Axel Boldt-Christmas]

On Fri, 20 Feb 2026 12:44:16 GMT, Ivan Walulya <iwalulya at openjdk.org> wrote:

>> src/hotspot/share/gc/serial/serialFullGC.cpp line 397:
>> 
>>> 395:     FlatArrayKlass* faklass = FlatArrayKlass::cast(array->klass());
>>> 396:     mark_and_push_closure.do_klass(faklass->element_klass());
>>> 397:   }
>> 
>> We have talked about this at the office a couple of times, but no sure if we reached any understanding.
>> 
>> But why do we feel the need for following the element klass in the flat array case. I understand conceptually that there are no headers with the element klass to visit here, but how is this different from an empty or all null reference array?
>> 
>> _PS: I see this code in Serial and G1, but not sure where to find it in Parallel if it exists._
>> 
>> _PPS: Also what about nested flattened values, (not sure if they occur with oops in the wild, yet) but it does not seem like we should rely on reaching the metadata this way. We should ensure that visiting the containers metadata will visit all its flattened fields/elements metadata._
>
>> PS: I see this code in Serial and G1, but not sure where to find it in Parallel if it exists.
> 
> I missed this, fixed now.
> 
> For the broader discussion on visiting element klass, in this PR, I just went with matching what happens with the case where we do not split arrays.

Yeah let us not address it in this RFE. But I think it something we need to consider, and evaluate what the actual requirements are when we trace through the metadata.

-------------

PR Review Comment: https://git.openjdk.org/valhalla/pull/2142#discussion_r2833038641